Created by Junhua Mao
This is a toolbox to download and manage the released part of the Pinterest40M multimodal dataset introduced in the paper Training and Evaluating Multimodal Word Embeddings with Large-scale Web Annotated Images. More information can be found on the [Project Page](http://www.stat. ucla.edu/~junhua.mao/multimodal_embedding.html).
If you find this dataset or toolbox useful in your research, please cite:
@inproceedings{mao2016training,
title={Training and Evaluating Multimodal Word Embeddings with Large-scale Web Annotated Images},
author={Mao, Junhua and Xu, Jiajing and Jing, Yushi and Yuille, Alan},
booktitle={NIPS},
year={2016}
}
Suppose that toolkit is install on $PATH_PTool:
cd $PATH_PTool
bash download_meta.shYou can easily download images in parallel (12 workers by default) and resize the downloaded images to 224x224:
cd $PATH_PTool
python download_images.pyThere are ~5 million images in the dataset. The download process can take days.
The script allows you to resume your downloading at any time. Just re-run download_images.py if your downloading is shutted down unexpectedly. It is possible that you failed to access some of the urls at the first time. Re-run download_images.py to have another try.
You are welcome to read download_images.py and py_utils.py for personalized and advanced downloading settings (e.g. see the docstring of py_utils.PinDataset.download_images).
View demo.ipynb for how to use this toolbox.
Use pin_2016_v1_0000.npy to pin_2016_v1_0097.npy as the training set.
Use pin_2016_v1_0098.npy as the validation set.
Use pin_2016_v1_0099.npy as the test set.
The copyright of the annotations and the images belongs to the original source. This meta data file can be used for research proposes only.
This toolbox is licensed under a Creative Commons Attribution 4.0 International License.